Note: Running pass on torch 1.7.1 and requirements in piano-transcription-master/pip.txt.
It seems there's some bug of miniconda on windows when using powershell7 and cmd. We may need to deactivate and reactivate the env. To ensure the correct installation, it's a good choice to check that the python lib paths of current env are included using sys in the terminal python.
#memo
python pytorch/inference.py --model_type Regress_onset_offset_frame_velocity_CRNN --checkpoint_path "300000_iterations.pth" --audio_path "C:/Users/Fred/Desktop/arr_April/speed_grapher/追忆.m4a" --cudaATEPP is a dataset of expressive piano performances by virtuoso pianists. The dataset contains 11742 11677 performances (~1000 hours) by 49 pianists and covers 1580 movements by 25 composers. All of the MIDI files in the dataset come from the piano transcription of existing audio recordings of piano performances. Scores in MusicXML format are also available for around half of the tracks. The dataset is organized and aligned by compositions and movements for comparative studies. More details are presented in the paper.
Please follow disclaimer.md to agree a disclaimer and download a latest version of ATEPP (~212MB).
You can inference your own track with the modified code and new checkpoint in piano_transcription-master. The env and setup are the same as https://github.com/bytedance/piano_transcription
python3 pytorch/inference.py --model_type=Regress_onset_offset_frame_velocity_CRNN --checkpoint_path=300000_iterations.pth --audio_path="resources/schumann_romanzen.mp3" --cuda
- 11742 performances (in midi format)
- 1007 hours
- 1580 movements
- 25 composers
- 49 performers
- 43% with scores
!Updates: 65 Pieces Removed
When creating ATEPP version-1.0, we only applied movement-wise matching to remove erroneously downloaded audio. Now, we finished detecting repeated audios by audio-wise fingerprint matching. Only 65 audios were detected repeated, and the corresponding transcribed midi files were removed. The repeats.csv lists the repeated transcribed files that have been removed.
Changed Statistics:
- 11677 performances
- 1002 hours
There are several issues found with the ATEPP Version-1.1:
-
Corrupted transcription results were found because of errors made by the transcription model when dealing with low-quality recordings. To detect the corrupted pieces, we compare the total note number and the duration of a performance with the medians of all the performances for the same composition. We filter out those which have similar duration to the midian but have much more/less notes with thresholds. Now ~60 pieces are confirmed to be corrupted and ~500 pieces are waiting for mannually checking. For thoes confirmed/suspected to be corrupted, we would not suggest to use them in the experiments.
-
Combined movements (the performance plays two or more movements) were found through a similar detection process of comparing the the total note number and the duration with the medians. ~170 pieces are suspected to have this issue.
-
Error labels of composition were found when we manually verify the suspected pieces.
Three csv files (combined_suspect.csv, corrupted_suspect.csv, corrupted_confirmed,csv) were uploaded. When using the dataset, you could exclude all pieces contained in these csv files. We will try to release a cleaner version soon. If you find any more issues or any errorous pieces that are not included in the above files, feel free to contact us.
We've released a Python package developed for linking classical music recording & track to the corresponding composition / movement, useful in cleaning up metadata in classical music datasets.
Package on PyPI: https://pypi.org/project/composition-entity-linker/
- Huan Zhang @github/anusfoil, huan.zhang@qmul.ac.uk,
- Jingjing Tang @github/BetsyTang, jingjing.tang@qmul.ac.uk
- Syed Rafee, @github/syedrafee s.rafee@qmul.ac.uk
@inproceedings{zhang2022atepp,
title={ATEPP: A Dataset of Automatically Transcribed Expressive Piano Performance},
author={Zhang, Huan and Tang, Jingjing and Rafee, Syed Rifat Mahmud and Fazekas, Simon Dixon Gy{\"o}rgy},
booktitle={ISMIR 2022 Hybrid Conference},
year={2022}
}
CC BY 4.0